AI Playground
Go ahead — try to run up my bill.
Most sites pay a fraction of a cent per message to somebody's API. This one doesn't. The model downloads into your browser and runs on your GPU, so you can talk to it until your laptop gets warm and it still costs this studio exactly nothing.
Pick a model. On first use it downloads into your browser, normally stays cached there, and then runs on your GPU — every message after that costs this studio exactly nothing.
Expect a real first-time download. The recommended model is about 295 MB, plus roughly 50 MB of shared retrieval/runtime files. The first load often takes several minutes; slower connections or devices can take longer. Larger choices take longer. Progress updates arrive in chunks, so the bar may pause and then jump; keep this tab open. Cached visits are faster but still need time to move the model into GPU memory.
Needs WebGPU · first load is a real model download · assets normally cache afterward · prompts stay on your device
What's actually happening
A whole RAG pipeline, with no server in it.
This is the same architecture we build for clients — retrieve, ground, generate, cite, refuse. The only unusual part is where it runs.
Embed
Your question is vectorised by all-MiniLM-L6-v2, in this tab.
Retrieve
Cosine similarity against a compact index built at deploy time.
Ground
The top chunks become the model's context, with a licence to refuse.
Generate
A local LLM streams the answer off your GPU, then cites what it used.
Reasonable questions
Where does this model actually run?
In your browser, on your GPU, via WebGPU. Model assets normally cache locally until the browser clears or evicts them, so later visits avoid the large download but still repeat GPU setup. Your prompts and generated answers are not sent to an inference server.
Does anything I type get sent anywhere?
No. Retrieval and generation both happen on your device. There is no endpoint to send it to.
Why is it a small model?
Because it has to fit in your browser. These are compact, browser-sized models — each is grounded in this site's content and told to refuse anything the content doesn't cover, which is what makes a small model useful rather than confidently wrong.
Is this what you'd build for a client?
The architecture, yes: retrieve, ground, generate, cite, refuse. The scale differs — a client system runs a larger model server-side or on managed infrastructure, over their real corpus, with evaluation.
Related
semantic rankingExplore the service path behind this topic.
- Related servicecosine match 0.63AI & LLM SolutionsClosest content match: AI & LLM Solutions.Explore service
- Related resourcecosine match 0.62Selected work: built, tested, measuredClosest content match: Selected work: built, tested, measured.Open resource
- Interactive democosine match 0.57The Lab — live demosClosest content match: The Lab — live demos.Open demo
Want this pointed at your documents?
The client version runs over your real corpus with evaluation, guardrails, and citations you can audit. Same shape, serious scale.